What happened
Pruna AI published P-H3-Image, an open-weights image model optimized from the MiniMax H3 video model through distillation and component rework, able to output similar people without extra training like LoRA.
Why it matters
A single reference image was enough to keep a character nearly consistent across new outfits and situations, a consistency previously associated with multimodal models such as NanoBanana and GPT-Image, though small text still breaks.
Summaries like this, every morning.
P-H3-Image arrives as an open-weights model, meaning the underlying weights are published for others to use, derived from a video model rather than built as an image model from the start. Pruna AI reached it by distillation and reworking components of MiniMax H3, so it carries over H3's ability to express a wide variety of images.
The practical draw is character consistency. In the test described, a single reference sheet combining a three-view drawing and expression variations, made about a year ago with NanoBanana, was enough to keep the character's look and clothing across coffee-drinking, bookstore and meal scenes. Adding character and clothing details to the prompt made the likeness take hold more reliably, and swapping outfits or situations remained straightforward.
The same test also surfaced the limit Pruna AI itself acknowledges: small text in generated images collapses. Readers weighing the model for character work may find it closes much of the gap with multimodal models like NanoBanana and GPT-Image on reference-image consistency, while text-heavy images remain a weak spot.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
